lkml.org 
[lkml]   [2006]   [Nov]   [29]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: Slab: Remove kmem_cache_t


On Wed, 29 Nov 2006, Nick Piggin wrote:
>
> You are saying that they should only be used to create new "primitive"
> types (ie. that you can use in arithmetic / logical ops) that can
> change depending on the config.

Well, it doesn't have to be something that is "arithmetic".

For an example of a primitive type that isn't arithmetic, the page table
entries are (pgt_t/pud_t/pmd_t/pte_t) are excellent - they don't do any
arithmetic or logical ops, but they do change depending on config, and no,
they aren't always opaque structures.

(Actually, these days they mostly are, but on many architectures it's much
slower to pass even a small struct around than it is to pass an integer
around - due simply to calling conventions - so for truly opaque things,
the typedef has the advantage that it _can_ be an opaque integer type, and
nobody will notice).

> That's fair enough. I'm sure you've also said in the past that they can
> be used (IIRC you even encouraged it) when the type is opaque in the
> context it is being used.

I'm sure I've been inconsistent, but in general, typedefs are bad. I think
you'll notice that I almost never use them myself. I much prefer passing
an opaque structure around, _unless_ I know the structure is so small that
it makes sense to do the above optimization (ie allow the case where the
opaque thing actually ends up being an integer).

Opaque integer types are generally useless in C, because they lose all the
type information _way_ too easily. There are no warnings for mis-use,
unless you use a sparse "bitwise" type and actually run sparse on the
thing. So even when there are performance reasons to use opaque integer
types (and on x86, the page table things were one such thing), usign a
struct is often preferable just for type-checking.

And as mentioned, there _are_ exceptions. Some types just get _sooo_
complex that it's inconvenient to type them out, even if they are
perfectly regular types, and don't depend on any config option. The
"filldir_t" typedef in fs.h is such an example - it's not really opaque,
_nor_ is it a config option, but it sure as hell would be inconvenient for
all low-level filesystems to do

int my_readdir(struct file *filp, void *dirent,
int (*filldir)(void *, const char *, int, loff_t,
u64, unsigned))
{
...
}

because let's face it, having to write out that "filldir" type just made
me use two lines (and potential for totally unnecessary tupos) because the
thing was so complex. So at that point, using a typedef is just common
sense, and we can do

int my_readdir(struct file *filp, void *dirent, filldir_t filldir)
{
...
}

instead.

But it's really quite hard to make that kind of complex type in C. It's
almost always a function pointer that takes complex arguments.

[ That said, I generally won't _complain_ if people use typedefs, but on
the other hand, some people definitely are too eager to do it, and I'll
happily remove them if people send me a patch. For example, we used to
have "task_t" for "struct task_struct", and that was just _unnecessary_,
and made it just harder to pick out what it was. Sometimes long names
and the explicit "struct" is a _good_ thing. ]

One final thing: for _small_ structures, typedefs are much better than for
large ones. Why? Because of stack usage. I want people to really _think_
about local variable sizes, and that's one thing that a typedef sometimes
causes - especially if it's opaque, so that users don't have any "handle"
on whether it is big or small, it's really nasty to use them for automatic
storage on the stack, because you may simply blow your stack usage on a
single (or a couple) of structures.

Making things be "struct something_or_other" makes at least _me_ think
more about it than if it's "file_t". Maybe it's just me, but I immediately
think "complex structure" when I see "struct", but "file_t" to me mentally
says "single word".

Linus
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2009-11-18 23:46    [W:0.031 / U:10.840 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site